HN survey: agent control planes (October 2026)

What Hacker News has seen, and upvoted, in agent orchestration, workbenches and control planes since January 2025.

Why I looked

I'm not after another way to run many agent sessions on one laptop; Claude Code already does that well. What I want is a control plane: a graph of work across many projects, where I plan, hand off to agents, see which agent is stuck, and get reminded of the context and what I need to do. Gas Town was the breakthrough here, because it put coordination on a shared ledger (fleet survey). This survey checks whether anyone has shipped that yet.

I searched HN (via Algolia) on 2026-10-30 for stories since January 2025 with more than 40 points. Points and comments as of today. Links go to the HN thread.

What's going on

• Running many sessions on a laptop is crowded and does fine. A desktop app or terminal that runs several Claude Code or Codex sessions in parallel, often as a kanban, typically gets 100–260 points. There are a dozen of these, and new ones keep coming.
• Sandboxed runtimes for agents are the hot topic right now. NVIDIA's OpenShell got 508 points on 22 Oct, and YC has funded several (Runtime, OneCLI). The question has moved from "how do I run ten agents" to "how do I let a team run them safely".
• Real orchestration posts do worse than you'd expect. Optio got 88 points and Hephaestus 81, and Paperclip never got past 4. The exceptions are first-party (Claude Agent Teams, 396) or a famous name (Gas Town, 354–403).
• The essays get the best discussion. Simon Willison on parallel agents, "your orchestrator is a bad clone of Elixir", and the Gas Town critiques all drew long threads. The recurring complaints: review is the bottleneck, costs run away, and handoff and memory between sessions are poor.
• Nobody owns the shared graph of work. Almost everything is one person's sessions on one machine. Hardly anything runs headless, holds the work for a whole team of people and agents, costs each task, and shows a person what needs them. That's the gap I keep running into (see "Why I looked"). Even Beads has competition: "I replaced Beads with a simpler Markdown tracker" got 84.

Many sessions, one laptop

Run agents remotely or in the cloud

Sandboxes and runtimes for teams

Orchestrators: breaking up work and handing it off

Company knowledge for agents

Essays and discussions

Related

• moc-agent-workbenches
• 2026-08-10-fleet-of-agents-survey — tool-by-tool notes on Paperclip, Gas Town, Optio and Agent Teams
• 2026-09-26-free-cursor-alternative-control-pane-for-ai — the laptop-side "control pane" question
• paperclip